讲解 Project 2: Classification Network辅导 Python编程

Project 2: Classification Network

Task1: CIFAR-10 Training

Dataset

CIFAR-10

Data distribution

Each team should select ONE of the following data distributions:

·IID Data Distribution

Ensure that the training dataset follows an IID (Independent and Identically Distributed)

pattern—i.e., an equal number of samples for each class.

i.e., 5000 images for each class.

·Non-IID Data Distribution

Use a non-IID training data setup. Specifically, the number of images in Classes 0–4 should be half the number in Classes 5–9.

i.e., 2500 images for each Classes 0–4; 2500*2 images for each Classes 5–9,

Challenge: 1-Minute Training on T4 GPU

Achieve at least 0.70 test accuracy within 1 minute of training using a T4 GPU.

Task2: Questions

Question 1 (Submission in the same .python file): Why does non-IID data distribution drive a

degrade in learning performance? What other factories could also result in performance degradation? Why?

Question 2 (Submission in the same .python file): What kind of neural network architecture(s)

is more suitable for learning in a limited time? (consider the challenge) Why?

Marking Scheme

Additional Notes

.    This is a group project. You are expected to complete all tasks collaboratively.

.    Submit only the `.ipynb` file (Jupyter Notebook or similar), which should be able to reproduce all your results.

.    Each group should decide whether to take on the IID or non-IID setup and adjust the data distribution accordingly.

.    In your submission: Clearly demonstrate that all performance criteria and challenges have been met. Include plots, explanations, and subtitles to make your submission easy to follow and understand.